1.5 KiB
Implementors must uphold all of the following for every call:
-
Memory domains: Pointers produced by to_ref() / to_mut() must be valid in the target execution domain for Self (e.g., CPU host memory for CPU, device memory for a specific GPU). If host↔device transfers are required, perform them inside the implementation; do not assume the caller synchronized.
-
Alignment & layout: All data must match the layout, stride, and element size expected by the kernel. size(), rows(), cols_in(), cols_out(), n(), etc... must be interpreted identically to the reference CPU implementation.
-
Scratch lifetime: Any scratch obtained from scratch.tmp_slice(...) (or a backend-specific variant) must remain valid for the duration of the call; it may be reused by the caller afterwards. Do not retain pointers past return.
-
Synchronization: The call must appear logically synchronous to the caller. If you enqueue asynchronous work (e.g., CUDA streams), you must ensure completion before returning or clearly document and implement a synchronization contract used by all backends consistently.
-
Aliasing & overlaps: If res, a, b, etc... alias or overlap in ways that violate your kernel’s requirements, you must either handle safely or reject with a defined error path (e.g., debug assert). Never trigger UB.
-
Numerical contract: For modular/integer arithmetic, results must be bit-exact to the specification. For floating-point, any permitted tolerance must be documented and consistent with the crate’s guarantees.