abstract class TensorBaseAll fields are embedded, with zero dynamic allocation. DType is a numeric enum, letting a switch over it compile to a jump table. Mirrors the C++ pattern: buffer + byteOffset + dtype.
The wrapper hierarchy exists to keep dtype and ownership explicit across addon and WASM boundaries.
TensorBasecarriesbuffer,byteOffset, anddtypeVectorBaseadds rank-1 semanticsVectorViewis a zero-copy borrowed wrapper over existing memoryVectorowns itsArrayBuffer
import { VectorView, Vector, DType, dot } from "numkong";
const raw = new Uint16Array([0x3c00, 0x4000, 0x4200]); // f16 payload, not uint16 values
const view = VectorView.from(raw, DType.F16);
const owned = new Vector(3, DType.F32);
owned.toTypedArray().set(new Float32Array([4, 5, 6]));
console.log(dot(view, view));
console.log(dot(owned, owned));
console.log(owned.byteLength);Use VectorView when the bytes already live somewhere else.
Use Vector when NumKong should own the storage.
Owned Vector and Matrix are fixed-capacity resizable. capacity reports the allocated element ceiling. tryResize(...) reshapes within capacity without moving the ArrayBuffer, so a toTypedArray() result stays valid. reserve(elements) may allocate a new ArrayBuffer to grow, which invalidates any TypedArray or row() handed out earlier. clear() empties the length while keeping capacity.
API
ClassTensorBaseAbstract base class for all tensor types.7
buffer
buffer: ArrayBufferbyteOffset
byteOffset: numberdtype
readonly dtype: DTypelength
readonly length: numberrank
readonly rank: numberbytesPerElementBytes per element for this tensor's dtype; the switch compiles to a jump table.
readonly bytesPerElement: numberbyteLengthTotal byte length of the tensor data.
readonly byteLength: numberClassVectorBaseAbstract rank-1 tensor base class.7
abstract class VectorBase extends TensorBaselength
length: numberrank
readonly rank: 1buffer
byteOffset
dtype
bytesPerElementBytes per element for this tensor's dtype; the switch compiles to a jump table.
byteLengthTotal byte length of the tensor data.
ClassVectorViewNon-owning rank-1 tensor view (like std::span<T>).10
class VectorView extends VectorBaseZero-copy wrapper for existing memory. Ideal for cross-module WASM interop where data already lives on the WASM heap.
VectorView
new VectorView(
buffer: ArrayBuffer,
byteOffset: number,
length: number,
dtype: DType,
)fromCreate a VectorView from any TypedArray, inferring or accepting dtype.
static from(arr: TypedArray, dtype?: DType): VectorViewto*× 3
toString(): stringlength
rank
buffer
byteOffset
dtype
bytesPerElementBytes per element for this tensor's dtype; the switch compiles to a jump table.
byteLengthTotal byte length of the tensor data.
ClassVectorOwning rank-1 tensor (like std::vector<T>).15
class Vector extends VectorBaseAllocates its own ArrayBuffer. Use for storing results or when you need independent ownership of the data.
Vector
new Vector(length: number, dtype: DType)new Vector(buffer: ArrayBuffer, length: number, dtype: DType)fromTypedArrayCreate an owning Vector by copying data from a TypedArray.
static fromTypedArray(arr: TypedArray, dtype?: DType): VectorfromViewCreate an owning Vector by copying data from any TensorBase.
static fromView(view: TensorBase): VectortoTypedArrayReturn a TypedArray view over this Vector's owned buffer, zero-copy.
toTypedArray(): TypedArraytryResizeResize within capacity without moving storage — toTypedArray() views stay valid.
tryResize(length: number): booleanReturns
false, length unchanged, if
length is negative or exceeds capacity().reserveGrow the allocated capacity to at least elements, reallocating and copying the live elements.
reserve(elements: number): booleanA no-op when already large enough; otherwise it invalidates any TypedArray already returned by toTypedArray(), which keeps pointing at the old buffer.
clearReset to length 0 while keeping the allocated capacity.
clear(): voidcapacityAllocated element capacity (>= length) — the ceiling tryResize honors.
readonly capacity: numberlength
rank
buffer
byteOffset
dtype
bytesPerElementBytes per element for this tensor's dtype; the switch compiles to a jump table.
byteLengthTotal byte length of the tensor data.
ClassMatrixBaseAbstract rank-2 tensor base class.11
abstract class MatrixBase extends TensorBaseAll 4 dimension fields are embedded — no dynamic allocation. Strides are in bytes to match the C API directly.
rows
rows: numbercolumns
columns: numberrowStride
rowStride: numbercolStride
colStride: numberlength
readonly length: numberrank
readonly rank: 2buffer
byteOffset
dtype
bytesPerElementBytes per element for this tensor's dtype; the switch compiles to a jump table.
byteLengthTotal byte length of the tensor data.
ClassMatrixOwning rank-2 tensor, row-major and C-contiguous by default.19
class Matrix extends MatrixBaseStrides are byte strides. Default for C-contiguous layout: rowStride = columns * bytesPerElement, colStride = bytesPerElement.
Matrix
new Matrix(rows: number, columns: number, dtype: DType)new Matrix(
buffer: ArrayBuffer,
byteOffset: number,
dtype: DType,
rows: number,
columns: number,
rowStride?: number,
colStride?: number,
)fromTypedArray
static fromTypedArray(
array: TypedArray,
rows: number,
columns: number,
dtype?: DType,
): MatrixtoTypedArray
toTypedArray(): TypedArrayrow
row(index: number): VectorViewtryResizeResize within capacity without moving storage (row()/toTypedArray() views stay valid), re-deriving C-contiguous strides.
tryResize(rows: number, columns: number): booleanReturns
false, unchanged, if either extent is negative or
rows*columns exceeds
capacity().reserveGrow the allocated capacity to at least elements, reallocating and copying the live rows*columns elements.
reserve(elements: number): booleanA no-op when already large enough; otherwise it invalidates any prior toTypedArray() / row() view. Intended for the C-contiguous owning layout.
clearReset to an empty 0x0 matrix while keeping the allocated capacity.
clear(): voidcapacityAllocated element capacity (>= rows*columns) — the ceiling tryResize honors.
readonly capacity: number