ascendc-st-design

Provide the operator's REQUIREMENTS.md and aclnn interface document, then follow the pipeline: define parameters, extract test factors with generate_test_factors.py, analyze constraints, generate implicit constraints and solver config, solve factor values, and finally run generate_test_cases.py for L0, L1, and L2 levels.

L0 cases are gate tests covering core functionality with single-factor coverage, capped at 200 cases. L1 cases use pairwise factor combination for functional, precision, and boundary testing, targeting 500-700 cases. L2 cases cover exception scenarios such as invalid dtype or dimension inputs, with at most 20 cases.

Nine constraint types are supported: calculate, broadcast_dim, broadcast_shape, conditional, match, existential, convertible, inferable_filter, and inferable. Each is defined in YAML with sources, target, and type-specific fields, and solved via topological ordering of the factor dependency graph.

Empty tensor cases are derived from existing normal L0/L1 cases rather than built from scratch. The script analyzes shape constraints to identify zero-able dimensions per operator type (matmul-like, reduce-like, or general), then modifies one template case per scenario so constraints remain satisfied.

All outputs go under operators/{operator_name}/tests/st/. Intermediate design artifacts (parameter definitions, factors, constraints, solver config, factor values) go in design/, while final L0/L1/L2 test case CSVs and coverage reports go in testcases/.