Expand description
Core tensor types, views, backend traits, and backend-independent contracts.
§Owned Tensors And Views
TypedTensor<T> and the dtype-erased Tensor enum are
owned tensor values. They are the right representation when a result is
materialized as compact column-major storage.
TypedTensorView is a borrowed typed view over an existing tensor buffer.
It carries logical shape, arbitrary strides, and an offset, so metadata-only
layout changes such as transposes, slices, and broadcasts can be represented
without copying. A view can be materialized explicitly with
TypedTensorView::to_contiguous when a compact owned tensor is required.
TensorRead is the dtype-erased borrowed input type used by eager kernels
and backend dispatch. It can borrow either an owned Tensor or a
TensorView with arbitrary strides. Prefer TensorRead for read-only
operation inputs so callers are not forced to materialize layout-only views.
TensorOwnedView and TensorValue are the owned lazy-value forms. Use
them when an API must store a view result beyond the lifetime of a borrowed
input, then expose a short-lived TensorRead at kernel-dispatch time.
Use Tensor::as_slice or TypedTensorView::as_slice only when compact
contiguous storage is part of the API contract. Use shape/stride-aware kernel
paths or TensorRead otherwise.
§Examples
use tenferro_tensor::{Tensor, TypedTensor};
let a = Tensor::F64(TypedTensor::from_vec_col_major(vec![2], vec![1.0, 2.0]).unwrap());
assert_eq!(a.shape(), &[2]);Re-exports§
pub use backend::default_backend_session;pub use backend::BackendCachedDot;pub use backend::BackendRuntimeCache;pub use backend::BackendSession;pub use backend::BackendSessionHost;pub use backend::SessionCachedDot;pub use backend::TensorAnalytic;pub use backend::TensorBackend;pub use backend::TensorBuffer;pub use backend::TensorDeviceTransfer;pub use backend::TensorDot;pub use backend::TensorElementwise;pub use backend::TensorFusion;pub use backend::TensorIndexing;pub use backend::TensorReduction;pub use backend::TensorStructural;pub use backend::TensorViewCanonicalization;pub use cache::CacheStats;pub use cache::RuntimeCacheControl;pub use config::*;pub use error::*;pub use types::*;
Modules§
- backend
- cache
- Cache accounting primitives shared by tensor backends and facade runtimes.
- config
- core
- Lightweight backend-independent host tensor data model.
- error
- Runtime error types for tensor execution.
- types
- validate
- Validation helpers shared across backends and exec layers.
Structs§
- Slice
Spec - Explicit slice descriptor.
Enums§
- Tensor
Ref - Core-neutral tensor input reference.