| name | performance-portability-review |
| description | Review Kokkos, MPI, GPU, or performance-portable work for correctness, parity, and maintainability. |
Performance Portability Review
When to Use
Use when reviewing or planning backend, GPU, Kokkos, MPI, or task-parallel implementation work.
Checklist
- Confirm reference correctness exists before optimization.
- Check memory layout, strides, ownership, and host/device boundaries.
- Inspect reductions, atomics, synchronization, and race conditions.
- Verify serial and parallel/backend parity tests.
- Check deterministic behavior where required.
- Confirm communication assumptions are documented.
- Require performance measurements only when the job asks for performance work.
- Avoid premature optimization before reference behavior is stable.
Output Format
Return:
- Correctness risks.
- Backend parity assessment.
- Memory and synchronization concerns.
- Test and benchmark assessment.
- Required follow-up.
Common Failure Modes
- Host/device copies are implicit or stale.
- Reductions are nondeterministic without documentation.
- MPI edge cases use one rank only.
- Optimized code diverges from the reference path.
- Benchmarks lack problem size, hardware, or build details.
