This directory contains Uni20's mdspan-facing structural utilities. It works with the Kokkos reference mdspan implementation through stdex:: and provides small helpers used by dense kernels and layout-aware algorithms.

mdspan.hpp: configured gateway to the Kokkos reference implementation in namespace stdex::.
mdspec.hpp: mdspan-shaped mapping and accessor metadata paired with a data descriptor for later handle acquisition.
format.hpp: std::format and {fmt} support for mdspan extents.
concepts.hpp: concepts for mdspan-like objects and related structural properties.
conjugate_accessor.hpp: read-only mdspan accessor adaptor and conj(...) view helper for lazy complex conjugation.
diagonal_accessor.hpp: generalized-diagonal logical mdspans backed by an explicit rank-one strided component view.
generated_accessor.hpp and generated_layout.hpp: read-only generated values with synthetic, non-strided logical offset mapping.
strides.hpp: stride inspection and stride utility helpers.
iteration_plan.hpp: backend-neutral compact affine iteration planning over one or more strided mappings, including element counts and reachable offset ranges.
transform_view.hpp: lazy read-only unary and variadic elementwise views.
zip_layout.hpp: helpers for matching or combining view layouts.

Notes

Include uni20/mdspan/mdspan.hpp when code needs the configured mdspan implementation itself.
This module should describe structure and layout, not ownership or backend dispatch policy. CPU execution of iteration plans belongs to the linalg CPU backend.
MdspanLike is the complete readable mdspan protocol used by leaf kernels: its descriptor aliases must agree, it exposes rank and extent observers, and its rank-dimensional operator[] returns the declared reference type. MutableMdspanLike additionally proves assignment through that indexed result.
MdspecLike is the broader structural protocol for mdspan metadata whose data handle is either immediately present or available through a data descriptor. mdspec is the standard unresolved materialization, but independent types may satisfy the concept directly. It preserves the actual mapping and accessor while intentionally exposing neither data_handle() nor element indexing.
MutableMdspecLike refines eventual write capability through an assignable accessor reference. It does not add indexing to an unresolved descriptor.
Immediate and descriptor-backed refinements use the same adjective order: mutable, ranked, strided, then the representation. For example, MutableRankedStridedMdspanLike requires an immediately usable mdspan, while MutableRankedStridedMdspecLike also accepts descriptor-backed metadata.
StridedMdspanLike refines MdspanLike by requiring an always-strided mapping and both mdspan and mapping stride observers. It includes layout_left, layout_right, and layout_stride mappings.
DiagonalMdspecLike refines MdspecLike without weakening its full logical rank-N interface. Off-diagonal access observes structural zero, while diagonal_components(span) exposes the physically stored rank-one values for specialized kernels. The component view preserves its signed stride, so a diagonal borrowed from another tensor need not be contiguous or forward ordered. The initial make_diagonal_mdspan factory accepts a default-accessor component mdspan and therefore produces a host-domain accessor. A future CUDA form must retain and propagate the CUDA component accessor rather than reclassifying a raw pointer-shaped handle.
An extent-one axis has no observable stride. Stride coalescing treats it as an identity: it never blocks a merge, and the merged descriptor retains the non-singleton axis's stride. Generic iteration plans remove singleton axes before coalescing.
A pointer data_handle_type is not enough to prove direct memory semantics. Backends that bypass access(...) must check for stdex::default_accessor or an explicitly lowerable accessor such as Uni20's conjugated_accessor.
Read-only accessor policies declare const element_type, including calculated accessors that return values. Ordinary MutableMdspanLike accessors also prove indexed assignment. An opaque accessor without assignable element semantics is not mutable merely because its handle is pointer-shaped.
uni20::const_access(span, indices...) performs read-only scalar access directly through a span's mapping and const-adapted accessor. Use it when a descriptor owner needs const element semantics without constructing a second mdspan for each access.
A tensor descriptor's const mdspan() overload resolves a const-element view. Mutable tensor access is exposed only by the non-const overload.
uni20::conj(span) is the user-facing lazy conjugation helper. Its accessor follows the C++26 std::linalg::conjugated_accessor direction while keeping Uni20's value-level conj behavior for real scalar types.
uni20::transform_view(function, spans...) constructs a read-only expression descriptor. A function evaluated through a CUDA-accessible transform accessor must make its invoked call operator device-callable with UNI20_HOST_DEVICE; CUDA kernel compilation validates that contract. Eager overwrite and update use the dispatched tensor operations assign_transform and transform_inplace.
Uni20-owned mapping, accessor, proxy-reference, generator, and transform execution surfaces are UNI20_HOST_DEVICE. This is independent of accessor domain registration: annotations make code compilable for a device, while enable_accessor_in_domain states where an acquired handle may be evaluated. Data descriptors and acquisition state remain backend-side control objects unless a leaf kernel explicitly defines a lowered descriptor ABI.
A device-callable mapping or accessor is not automatically supported by every precompiled CUDA backend. Such a backend must carry a sufficient type-erased plan, provide an explicitly instantiated typed lowering, or cleanly decline.

Contents

Notes

Related Documentation