Expand description
Hardware abstraction layer (docs/SPEC.md §2.2): the object-safe GpuBackend
trait, opaque backend-tagged handles, the always-available CpuBackend
(the correctness reference), the opt-in CUDA and Metal backends, and the
HardwareDetector that picks one at runtime.
GPU backends are cargo features (cuda, metal) and never part of the
default build; metal is additionally macOS-only.
Dependency direction: depends on oxidelake-core and oxidelake-memory.
Re-exports§
pub use args::AggregateArgs;pub use args::FilterProjectArgs;pub use args::HashJoinArgs;pub use args::JoinBuildArgs;pub use args::JoinProbeArgs;pub use args::VectorDistanceArgs;pub use backend::GpuBackend;pub use backend::MemoryInfo;pub use cpu::CpuBackend;pub use detector::Availability;pub use detector::BACKEND_ENV;pub use detector::HardwareDetector;pub use handles::DeviceBatch;pub use handles::DeviceBuffer;pub use handles::DeviceJoinMatches;pub use handles::DeviceJoinTable;pub use handles::DeviceStream;
Modules§
- args
- Argument bundles for the operator entry points of
crate::GpuBackend. - backend
- The object-safe hardware abstraction.
- columns
- Byte-level (de)serialization of Arrow columns for device transfers.
- cpu
- The CPU backend: always available, fully functional, and the semantic reference every GPU backend must match.
- cuda
- The CUDA backend (feature
cuda). - detector
- Runtime backend selection: probe CUDA → Metal → CPU, honouring
OXIDE_BACKEND. - handles
- Opaque, backend-tagged handles. Their internals are enums over the enabled
backends; every handle frees its resource on
Dropthrough the owning backend’s RAII types.
Structs§
- Device
Id - Identifies one device of a backend, e.g.
cuda:0. - Host
Buffer - A host buffer whose memory class (pinned/pageable) is observable. A host buffer whose memory class is observable.
Enums§
- Backend
Kind - The execution backends OxideLake knows about.