compute.GPUndarray.empty¶
Summary¶
compute.GPUndarray.empty allocates a GPU-resident ndarray backed by a storage buffer. You provide dtype and layout metadata; the buffer is allocated but not initialized. Use this for output tensors, scratch arrays, or GPU-only workflows. Descriptor options map to storage-buffer creation flags.
Syntax¶
WasmGPU.compute.GPUndarray.empty(ctx: { device: GPUDevice; queue: GPUQueue; readback?: ReadbackRing }, dtype: DType, layout: NdLayoutDescriptor, desc?: Omit<StorageBufferDescriptor, "byteLength" | "data">): GPUndarray
const g = wgpu.compute.GPUndarray.empty(wgpu.compute, dtype, layout, desc);
Parameters¶
| Name | Type | Required | Description |
|---|---|---|---|
ctx |
{ device: GPUDevice; queue: GPUQueue; readback?: ReadbackRing } |
Yes | GPU context used to allocate backing storage. Passing wgpu.compute also associates its shared readback ring with the result. |
dtype |
DType |
Yes | Element data type. |
layout |
NdLayoutDescriptor |
Yes | Shape plus optional byte strides and byte offset. |
desc |
Omit<StorageBufferDescriptor, "byteLength" \| "data"> |
No | Optional storage-buffer descriptor overrides. |
Returns¶
GPUndarray - Newly allocated GPU ndarray.
Example¶
const canvas = document.querySelector("canvas");
const wgpu = await WasmGPU.create(canvas);
const g = wgpu.compute.GPUndarray.empty(wgpu.compute, "f32", { shape: [128, 128] }, { copySrc: true });
console.log(g.residency, g.layout());