Crate rstsr_core 

API document of rstsr-core

rstsr-core is the core of RSTSR series:

• It defines data structure and some traits (interface) of tensor, storage, device.
• It realizes two basic devices: DeviceCpuSerial and DeviceFaer, so rstsr-core alone is a functional tensor toolkit library.

If you are more aware of matmul efficiency, or by other considerations (we will try to implement BLAS and Lapack features in future), you may find DeviceOpenBLAS in rstsr-openblas helpful. We hope to implement more devices in future.

User document

User document refers to readthedocs. This document is still in construction.

Cheat sheet

For NumPy/SciPy users, the rstsr facade crate ships a cheat sheet mapping common operations to their RSTSR equivalents:

• the core operations page — printing and layout, creation, indexing, manipulation, operators, reductions, sorting;
• the linear algebra page — the faer-backed rt::linalg family.

Each row is verified by an integration test (rstsr/tests/cheatsheet_*.rs) and by a NumPy/SciPy parity script; the authoring conventions live in the skill api-cheatsheet-conventions.

API specifications

API specifications is provided for summary of important identifiers (structs, traits and functions) in rstsr-core.

Currently, more detailed documentation for each identifiers are still constructing.

We also provide fulfillment for Array API standard. This fulfillment check shows how much functionalities we have achieved to be a basic tensor (array) library.

For NumPy users, in this meantime, fulfillment for Array API standard shows the similar parts of RSTSR and NumPy; and these parts in API specifications shows the difference between RSTSR and NumPy:

• Tensor Structure and Ownership
• RSTSR Specific Identifiers
• Storage-dependent manipulation
• Developer Area

Variable naming convention

Abbreviations of type annotations

Abbreviation	Meaning
TR	Tensor type (struct TensorBase)
T	Data type (can be f64, Complex<f32>, etc)
B	Backend (device) type (applies DeviceAPI)
D	Dimensional (applies DimAPI)
S	Storage (struct storage::exports::Storage)
R	Data with lifetime and mutability (applies DataCloneAPI)
C	Content (usually Vec<T> for CPU devices)

Variable or Names

Rule	Convention
suffix API	Traits. To avoid naming collision with structs or other crates. Example: DeviceAPI, AsArrayAPI
suffix _f	Fallible functions. Those functions pairs with a function that panics when error. Example: zeros_f, sin_f, reshape_f
prefix ext_	Functions that fulfill the Array API mixed-dtype promotion contract: the operands may have different data types, and the result is computed in their promoted common data type (the same rule as NumPy). The same-dtype functions instead require matching operand data types. Example: ext_add, ext_matmul, ext_vecdot, ext_outer
prefix to_ or no prefix (manipulation functions)	Ensures that input is reference. Example: reshape, transpose, to_dim
prefix into_ (some manipulation functions)	Only changes layout, and does not change data with its lifetime and mutability. Example: into_broadcast, into_transpose, into_dyn
prefix into_ output is always Tensor (some manipulation functions)	Give owned tensor. This special case will occur when there is possiblity to generate a new tensor by manipulation function. For these cases, prefix to_ and change_ will give output TensorCow, prefix into_ will give Tensor. Example: into_shape, into_layout
prefix change_ output is always TensorCow (some manipulation functions)	Give copy-on-write tensor by consuming input. This special case will occur when there is possiblity to generate a new tensor by manipulation function. Example: change_shape, change_layout
type annotation Args	Input tuple type as overloadable parameters. Example: zeros, asarray
type annotation Inp	Additional annotation that rust requires for type deduction. Example: zeros, asarray

Re-exports

pub use tensorbase::Tensor;

pub use tensorbase::TensorAny;

pub use tensorbase::TensorArc;

pub use tensorbase::TensorBase;

pub use tensorbase::TensorCow;

pub use tensorbase::TensorMut;

pub use tensorbase::TensorRef;

pub use tensorbase::TensorReference;

pub use tensorbase::TensorView;

pub use tensorbase::TensorViewMut;

Modules

api_specification

API Specification of crate rstsr-core

array_api_standard

Array API standard Fulfillment

dev_utilities 🔒

device_cpu_serial

Backend for CPU, serial only.

device_faer

Backend for CPU, some using rayon for parallel, but matmul and linalg implemented by faer.

feature_rayon

This module contains some features for rayon (parallel).

format

Tensor formatting: Display and Debug implementations for tensor types.

operators

Operations on tensors.

order_semantics

Device default orders: RowMajor and ColMajor

prelude

User-facing prelude of rstsr-core.

prelude_dev

Developer prelude of rstsr-core: convenience re-exports for writing device implementations, operators, and tests against rstsr-core internals (core/alloc utilities, macro tools, and the exports of internal crates). User code should prefer crate::prelude instead.

storage

Basic storage type and trait definitions.

tensor

Tensor functions organized by category: creation, indexing, manipulation, element-wise operators, reduction, and linalg, plus ownership, device, and external-type conversions.

tensorbase

Fundamental tensor types: TensorBase and the ownership aliases specialized from it (Tensor, TensorView, TensorMut, TensorCow, TensorArc), plus the fully generic alias TensorAny.

Macros

allclose

tensor_from_nested

Create a tensor from a nested array literal (for development purposes).

Type Aliases

DeviceCpu