Tensor API

GPU tensor with shape and layout metadata.

---

Constructor

constructor(context: GPUContext, shape: TensorShape, options?: TensorOptions)

Create a new tensor with uninitialized data.

Parameters:

• context: GPU context
• shape: Tensor dimensions
• options: Configuration options

interface TensorOptions {
  layout?: DataLayout;  // Default: 'NCHW'
  label?: string;
}

---

Static Methods

zeros

static zeros(context: GPUContext, shape: TensorShape, options?: TensorOptions): Tensor

Create a tensor filled with zeros.

Example:

const tensor = Tensor.zeros(context, [1, 3, 224, 224]);

---

ones

static ones(context: GPUContext, shape: TensorShape, options?: TensorOptions): Tensor

Create a tensor filled with ones.

---

filled

static filled(context: GPUContext, shape: TensorShape, value: number, options?: TensorOptions): Tensor

Create a tensor filled with a specific value.

---

fromArray

static fromArray(
  context: GPUContext,
  data: number[] | Float32Array,
  shape: TensorShape,
  options?: TensorOptions
): Tensor

Create a tensor from an array of numbers.

Example:

const data = [1, 2, 3, 4, 5, 6];
const tensor = Tensor.fromArray(context, data, [2, 3]);

---

Methods

upload

async upload(data: Float32Array): Promise<void>

Upload data to the GPU buffer.

Parameters:

• data: Data to upload (must match tensor size)

Throws:

• BufferSizeError: If data length doesn't match tensor size

---

download

async download(): Promise<Float32Array>

Download data from the GPU buffer.

Returns: Float32Array containing tensor data

Example:

const tensor = Tensor.zeros(context, [2, 3]);
await tensor.upload(new Float32Array([1, 2, 3, 4, 5, 6]));
const data = await tensor.download();
console.log(data); // Float32Array [1, 2, 3, 4, 5, 6]

---

reshape

reshape(newShape: TensorShape): Tensor

Create a view with a new shape. Zero-copy operation.

Parameters:

• newShape: New dimensions (total elements must match)

Returns: New tensor view sharing the same GPU buffer

Example:

const tensor = Tensor.zeros(context, [1, 3, 224, 224]);
const view = tensor.reshape([1, 150528]); // Zero-copy

---

convertLayout

async convertLayout(target: DataLayout): Promise<Tensor>

Convert memory layout (NCHW ↔ NHWC).

Parameters:

• target: Target layout

Returns: New tensor with converted layout

Example:

const nchw = Tensor.zeros(context, [1, 3, 224, 224], { layout: 'NCHW' });
const nhwc = await nchw.convertLayout('NHWC');
console.log(nhwc.shape); // [1, 224, 224, 3]

---

destroy

destroy(): void

Release GPU buffer. No-op for view tensors.

Note: Only the original tensor that owns the buffer frees GPU memory.

---

usesContext

usesContext(context: GPUContext): boolean

Check if this tensor was created with the given context.

---

Properties

shape

readonly shape: TensorShape

Tensor dimensions.

---

layout

readonly layout: DataLayout

Memory layout (NCHW or NHWC).

---

buffer

readonly buffer: GPUBuffer

Underlying GPU buffer.

---

size

readonly size: number

Total number of elements.

---

ownsBuffer

readonly ownsBuffer: boolean

Whether this tensor owns the GPU buffer (false for views).

---

Memory Management

View Pattern

// Original tensor owns the buffer
const original = new Tensor(context, [1, 3, 224, 224]);
original.ownsBuffer;  // true

// View shares the buffer (zero overhead)
const view = original.reshape([1, 3 * 224 * 224]);
view.ownsBuffer;  // false

// Cleanup
view.destroy();   // No-op: doesn't own buffer
original.destroy();  // Actually destroys GPU buffer

Data Roundtrip

const data = new Float32Array([1, 2, 3, 4, 5, 6]);
const tensor = Tensor.fromArray(context, data, [2, 3]);
const result = await tensor.download();
// result equals data