Kernel Validation

The validation sequence distinguishes structural validity, numerical correctness, hardware behavior, and static schedule data.

1. Kernel compilation (cargo furiosa-opt compile) translates and verifies every #[device] body, catching mapping and shape errors without executing anything.
2. CPU runs (plain cargo) compare host-side tensor results with a host oracle.
3. NPU execution runs the compiled executable device format (EDF) through the Furiosa SDK and physical hardware.
4. Schedule data inspects emitted schedule JSON and compares makespan with release, shapes, types, and mappings fixed.

Kernel compilation is static: it proves that every selected kernel translates and verifies, without providing numerical-correctness proof. Compare schedule makespan only when the release, shapes, data types, and mapping are fixed. A schedule is not a throughput measurement by itself. Use Scheduling and Tuning to compare fixed schedule candidates. Record the source revision, backend, shapes, mappings, oracle result, and schedule artifact for each comparison.