Struct ErasedHostTensor 

pub struct ErasedHostTensor { /* private fields */ }

A host tensor whose element type is recovered at run time.

The payload keeps its own concrete type and is recovered by identity, so no byte reinterpretation happens and a caller-owned payload is duplicated through its own entry point rather than by copying bytes.

The value also carries the layout of the view it presents. Clone shares that payload, so a metadata-only permutation is cheap; duplicate copies it, so a caller that needs its own storage asks for one explicitly. A typed read applies the layout, and the dense accessors refuse a strided view instead of presenting the payload as if it were the view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let value = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2], vec![1.0_f64, 2.0])?);
assert_eq!(value.downcast_ref::<f64>().unwrap().as_slice(), &[1.0, 2.0]);
assert_eq!(value.clone().element_count(), 2);

Implementations

impl ErasedHostTensor

pub fn new<T: Scalar>(value: TypedTensor<T, DynRank, Host>) -> Self

Erase a host tensor’s element type.

The result presents the tensor’s dense column-major layout.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![7_i32])?);
assert!(erased.is::<i32>());

pub fn type_id(&self) -> TypeId

Identity of the stored element type.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![1.0_f32])?);
assert_eq!(erased.type_id(), core::any::TypeId::of::<TypedTensor<f32, DynRank, Host>>());

pub fn element_type_id(&self) -> TypeId

Identity of the stored element type, without the tensor wrapper.

This is what a runtime tag reports for an externally defined scalar.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![1.0_f64])?);
assert_eq!(erased.element_type_id(), core::any::TypeId::of::<f64>());

pub fn shape(&self) -> &[usize]

Shape of the presented view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 3], vec![0.0_f64; 6])?);
assert_eq!(erased.shape(), &[2, 3]);

pub fn strides(&self) -> &[isize]

Element strides of the presented view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 3], vec![0.0_f64; 6])?);
assert_eq!(erased.strides(), &[1, 2]);

pub fn offset(&self) -> isize

Element offset of the presented view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2], vec![0.0_f64; 2])?);
assert_eq!(erased.offset(), 0);

pub fn is_contiguous(&self) -> bool

Whether the presented view is the dense column-major layout of its shape.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 3], vec![0.0_f64; 6])?);
assert!(erased.is_contiguous());

pub fn element_count(&self) -> usize

Number of elements in the presented view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 3], vec![0.0_f64; 6])?);
assert_eq!(erased.element_count(), 6);

pub fn shares_payload_with(&self, other: &Self) -> bool

Whether two erased values present the same stored payload.

A metadata-only view answers true, and an independent copy answers false.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![1.0_f64])?);
assert!(erased.shares_payload_with(&erased.clone()));
assert!(!erased.shares_payload_with(&erased.duplicate()));

pub fn duplicate(&self) -> Self

Copy the payload into an independent value with the same view.

The copy is the caller’s own storage, so later mutation of either value leaves the other unchanged.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![1.0_f64])?);
let mut copy = erased.duplicate();
copy.downcast_mut::<f64>().unwrap().host_data_mut()[0] = 5.0;
assert_eq!(erased.as_dense::<f64>().unwrap().0, &[1.0]);

pub fn permuted(&self, axes: &[usize]) -> Result<Self>

Present the same elements under a permuted axis order.

This is metadata only: the payload is shared, no element is moved, and the result’s shape and strides follow axes.

Errors

Returns a validation error carrying tenferro_tensor_core::ValidationError::InvalidPermutationLength when axes does not have one entry per axis, tenferro_tensor_core::ValidationError::AxisOutOfBounds when an axis is out of range, or tenferro_tensor_core::ValidationError::DuplicateAxis when an axis repeats.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 3], vec![0.0_f64; 6])?);
let permuted = erased.permuted(&[1, 0])?;
assert_eq!(permuted.shape(), &[3, 2]);
assert_eq!(permuted.strides(), &[2, 1]);
assert!(permuted.shares_payload_with(&erased));

pub fn to_contiguous(&self) -> Result<Self>

Materialize the presented view into a dense column-major payload.

The result owns its elements in the view’s axis order, so a subsequent read with the dense accessors returns the same logical values in a contiguous buffer.

Errors

Returns a validation error carrying tenferro_tensor_core::ValidationError::ShapeDataLengthMismatch when the view names storage the payload does not have.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 2], vec![1.0_f64, 2.0, 3.0, 4.0])?);
let contiguous = erased.permuted(&[1, 0])?.to_contiguous()?;
assert_eq!(contiguous.shape(), &[2, 2]);
assert!(contiguous.is_contiguous());
assert_eq!(contiguous.as_dense::<f64>().unwrap().0, &[1.0, 3.0, 2.0, 4.0]);

pub fn is<T: Scalar>(&self) -> bool

Whether the stored tensor has element type T.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![1.0_f64])?);
assert!(erased.is::<f64>());

pub fn downcast_ref<T: Scalar>(&self) -> Option<&TypedTensor<T, DynRank, Host>>

Borrow the whole payload when it is dense and has element type T.

A strided view answers None rather than presenting the payload as if it were the view. Use ErasedHostTensor::element_at or ErasedHostTensor::to_contiguous for a view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 2], vec![1.0_f64, 2.0, 3.0, 4.0])?);
assert_eq!(erased.downcast_ref::<f64>().unwrap().shape(), &[2, 2]);

// A strided view is not the dense payload, so the dense borrow refuses it.
assert!(erased.permuted(&[1, 0])?.downcast_ref::<f64>().is_none());

pub fn downcast_mut<T: Scalar>( &mut self, ) -> Option<&mut TypedTensor<T, DynRank, Host>>

Mutably borrow the whole payload when it is dense, unique, and has element type T.

A strided view or a payload shared with another value answers None, so aliasing is never reachable through this entry point. Use ErasedHostTensor::element_at_mut for a view.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let mut erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![1_i64])?);
erased.downcast_mut::<i64>().unwrap().host_data_mut()[0] = 9;
assert_eq!(erased.downcast_ref::<i64>().unwrap().as_slice(), &[9]);

pub fn into_typed<T: Scalar>(self) -> Option<TypedTensor<T, DynRank, Host>>

Take the whole payload when it is dense and has element type T.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![2.0_f64])?);
assert_eq!(erased.into_typed::<f64>().unwrap().as_slice(), &[2.0]);

pub fn payload_element_count(&self) -> usize

Number of elements the stored payload holds.

A strided view may reach fewer than all of them, so this is the extent a buffer-length check must use rather than the view’s own count.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 2], vec![0.0_f64; 4])?);
assert_eq!(erased.payload_element_count(), 4);

pub fn as_dense<T: Scalar>(&self) -> Option<(&[T], &[usize])>

Borrow the dense element slice and its shape when the view is contiguous and has element type T.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2], vec![1.0_f64, 2.0])?);
assert_eq!(erased.as_dense::<f64>().unwrap().0, &[1.0, 2.0]);

pub fn element_at<T: Scalar>(&self, index: &[usize]) -> Option<&T>

Borrow one element of the presented view by logical index.

This applies the view’s strides and offset, so it reads the element the view names even when the view is not contiguous.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 2], vec![1.0_f64, 2.0, 3.0, 4.0])?);
let permuted = erased.permuted(&[1, 0])?;
assert_eq!(permuted.element_at::<f64>(&[1, 0]), Some(&3.0));
assert_eq!(permuted.element_at::<f64>(&[2, 0]), None);

pub fn element_at_mut<T: Scalar>(&mut self, index: &[usize]) -> Option<&mut T>

Mutably borrow one element of the presented view by logical index.

This applies the view’s strides and offset. It answers None when the payload is shared with another value, so two live views never produce two mutable borrows of one element. A writer therefore holds the only reference: build the view, release the value it came from, and mutate through the view, or call ErasedHostTensor::duplicate for a payload of its own.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let erased = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![2, 2], vec![1.0_f64, 2.0, 3.0, 4.0])?);

// A view that shares its payload refuses a mutable element borrow.
let mut shared = erased.permuted(&[1, 0])?;
assert!(shared.element_at_mut::<f64>(&[1, 0]).is_none());

// An independent copy accepts it, and the original stays unchanged.
let mut owned = erased.permuted(&[1, 0])?.duplicate();
*owned.element_at_mut::<f64>(&[1, 0]).unwrap() = 20.0;
assert_eq!(owned.element_at::<f64>(&[1, 0]), Some(&20.0));
assert_eq!(erased.as_dense::<f64>().unwrap().0, &[1.0, 2.0, 3.0, 4.0]);

Trait Implementations

impl Clone for ErasedHostTensor

fn clone(&self) -> Self

Share the payload and copy only the layout metadata.

This is a metadata-only operation, so it is what a permutation and a metadata view use. Use ErasedHostTensor::duplicate for an independent copy of the storage.

Examples

use tenferro_tensor::{DynRank, ErasedHostTensor, Host, TypedTensor};

let value = ErasedHostTensor::new(TypedTensor::<_, DynRank, Host>::from_host_vec_col_major(vec![1], vec![7_i64])?);
let view = value.clone();
assert!(view.shares_payload_with(&value));
assert!(!view.duplicate().shares_payload_with(&value));

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ErasedHostTensor

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.