Host/device portability tiers

Slam’s types are designed to run both on the host and inside GPU kernels.

Not every C++ facility is safe in device code across all of Slam’s backends (sequential, OpenMP, CUDA, HIP), so Slam categorizes the constructs it uses into three portability tiers. Compile-time features generate no device code and are therefore unconditionally kernel-safe.

Table 14 Slam’s portability tiers

Tier	Contents	Allowed where
A	All compile-time language features (concepts, if constexpr, class template argument deduction (CTAD), non-type template parameters, fold expressions, type traits, constexpr evaluation) as well as Axom host-device types (StackArray, ArrayView, NumericLimits, utilities::*).	everywhere, including kernels
B	std::string_view, std::variant, std::ranges views/algorithms, std::vector, std::map, exceptions, and iostreams.	host only: builders, registries, isValid(verbose), and I/O
C	Virtual functions on host-device types; standard containers held inside view types; throwing accessors on host-device paths.	nowhere (existing instances are migration targets)

Device std::optional

Slam uses std::optional for optional-returning APIs. The CUDA host-configs enable --expt-relaxed-constexpr, and HIP’s compiler accepts these standard library calls from host-device paths in the supported builds.

The device-side contract is to check has_value() before dereferencing with operator*, or to use value_or. Avoid value() in kernels because standard library implementations can route the disengaged case through a throwing host-only helper.